The Seasonality of Con�ict∗

Jenny Guardado† and Steven Pennings‡

This Draft: July 2017

Abstract

This paper exploits the seasonality of agricultural labor markets to estimate the

e�ect of changes in the returns to working on con�ict intensity. Using a dynamic model

of labor supply, we �rst show theoretically that exogenous, anticipated, and transitory

changes in labor demand due to harvest are better able to capture the e�ects of changes

in the opportunity cost of con�ict relative to other shocks commonly analyzed in the

literature. This is because harvest shocks hold constant other dynamic drivers of con-

�ict which can bias empirical estimates and obscure the strength of the opportunity

cost mechanism. Our empirical identi�cation strategy exploits exogenous sub-national

variation in the timing and intensity of harvest driven by local climatic conditions. Us-

ing data from several con�ict settings � Afghanistan, Iraq, and Pakistan � our results

show that the onset of harvest usually leads to a statistically signi�cant reduction in

the share of monthly insurgent attacks.

∗We are grateful for comments from Scott Ashworth, Eli Berman, Ethan Bueno de Mesquita, James Fearon, Anthony

Fowler, Noel Maurer, Matias Iaryczower, John Londregan, and participants of the 2017 Cesifo Venice Summer Institute, the

2017 Princeton Political Economy Seminar; the 2015 Third Formal & Comparative Conference (Becker-Friedman Institute);

World Bank MENA Seminar Series; Harris School of Public Policy Political Economy Lunch; the 2015 HiCN (Households in

Con�ict Network) conference; the Georgetown Political Economy & Development Seminars; and the 2016 Warwick-Princeton

Political Economy Workshop. We thank �nancial support from Ethan Bueno de Mesquita and the O�ce of Naval Research;

and are grateful to Paula Ganga, Jeong Whan Park, Marissa Barragan and Saisha Mediratta for their research assistance.†Assistant Professor. Georgetown University. jennyguardado@gmail.com. Project developed while visiting the Harris School

of Public Policy University of Chicago (2014-2015).‡Research Economist. World Bank Research Department spennings@worldbank.org. The views expressed here are the

authors', and do not necessarily re�ect those of the World Bank, its Executive Directors, or the countries they represent.

1

1 Introduction

Understanding why civil con�ict takes place is of utmost policy importance for the developing

world given its toll on human lives, human capital, and broader development prospects. The

current view among international institutions such as the World Bank, academics, and even

public opinion1 is that economic factors such as poverty and unemployment are among the

main drivers of con�ict. Indeed, a number of theoretical (Becker 1968; Grossman 1991; Dal

Bo and Dal Bo 2011) and empirical studies establish a connection between the returns to

working and the intensity of con�ict � implying that better wages or job prospects increase

the opportunity cost of �ghting versus working. Typically, these studies show that negative

income shocks driven by rainfall (Miguel et. al. 2004) or commodity prices (Dube and

Vargas 2013; Guardado 2016; Hodler and Raschky 2014, among others) are associated with

an increase in the likelihood and intensity of civil con�lict. Yet, despite this evidence, other

studies question both the strength and interpretation of the relationship between economic

factors and con�ict (Blattman and Bazzi 2014; Berman et. al. 2011b) and whether economic

considerations are an important driver of civilian participation in con�ict altogether (Berman

et. al. 2011a).

In this paper, we make two arguments: �rst, we show that these con�icting empirical

�ndings on the opportunity cost mechanism are likely driven by the type of income shocks

commonly analyzed in the literature. We demonstrate that when income shocks are highly

persistent it leads to empirical results which systematically underestimate the strength of

the opportunity cost mechanism. This is the case of income shocks driven by commodity

prices � which tend to be persistent. Indeed, simulation data from di�erent theoretical

models shows that regressions estimates of the opportunity cost mechanism are systemat-

ically upward biased and could range from the negative to the positive depending on the

degree of persistence. This range of estimates is consistent with the mixed empirical results

in the literature and the ambiguity over the impact of income shocks documented in land-

mark models of con�ict (Fearon 2008). These �ndings also suggest that the opportunity cost

e�ect captured by commodity prices is likely to be even stronger (or more negative) than

shown in current studies (e.g. Dube and Vargas 2013; Blattman and Bazzi 2014; Guardado

2016; Hodler and Raschky 2014, etc.).

1See WDR 2011.

2

Given these estimation concerns from examining income shocks, we propose a di�erent

way to empirically gauge the strength of the opportunity cost mechanism: by exploiting

seasonal changes in labor demand driven by the timing and intensity of harvest. Due to

the temporary and anticipated nature of harvest, we are able to hold constant important

non-wage determinants of con�ict such as the value of winning (Fearon 2008, Chassang and

Padro-i-Miquel 2009) or the marginal utility of consumption, which normally upward biases

empirical results. Indeed, estimates using simulated data show that the �true� opportunity

cost of con�ict can be uncovered almost exactly by a regression of time allocated to violence

on seasonal variation in wages.

We focus our attention on the e�ect of seasonal labor demand shocks in Iraq (2004-2009),

Pakistan (1988-2010), and Afghanistan (2004-2007), driven by the wheat harvest calendar

(the main legal agricultural crop). Given the labor-intensive nature of agricultural activities,

and the fact that many of these crops are harvested annually, they induce a large, transitory

and anticipated change in the local demand for labor. Since the timing and intensity of

harvest is determined by local climatic conditions (which we measure using pre-con�ict

data), we can rule out reverse causality running from con�ict to opportunity costs. Because

monthly time-series for local wages are normally lacking, we focus on the reduced-form

relationship between the number of attacks in a location and the size of the area harvested.

Estimates across di�erent con�ict settings (Iraq, Pakistan, and Afghanistan) show that

at times of greater labor demand due to harvest, the intensity of con�ict is lower compared

to non-harvest times in districts with a smaller area harvested. Our main �ndings show that

for a district with the average crop intensity, the onset of harvest reduces the average share

of monthly attacks by around 25% in Pakistan, 11-13% in Iraq, and 22% in Afghanistan. In

addition, using household surveys and monthly weather information, we are able to rule out

alternative explanations based on temperature, precipitation, state-driven violence, religious

calendars, seasonal migration or job switching. Instead, consistent with our interpretation we

show that during harvesting months agricultural workers tend to have di�erentially higher

employment rates relative to other rural workers. Overall, these results indicate that in

di�erent con�ict settings the opportunity costs of �ghting � the foregone returns from

working � may play a key role in determining the intensity of con�ict.

Additional qualitative evidence suggests that such a trade-o� between working and �ght-

ing is particularly applicable to �part-time� �ghters � individuals who shift between con�ict

and legal work � depending on changing economic opportunities. In fact, part-time �ghters

3

are a common feature of the industrial organization of modern insurgencies.2 For example,

it is well known that some of members of the Vietcong guerrilla worked as farmers during

the day but fought US forces at night. Con�ict in the Philippines also explicitly relied on

part-time �ghters during the 1990s when entire battalions from the Moro Islamic Liberation

Front (MILF) employed part-time soldiers on a monthly rotational basis to aid full-time com-

batants (Cline 2000). Similarly, in Afghanistan, Taliban forces have been known to organize

in village cells each containing around ten to �fty part-time �ghters (Afsar et. al. 2008).

This was also the case for Iraq during the US intervention3, as well as of highly ideological

guerrillas such as Shining Path in Peru in the 1980s (McClintock 1998). The reason for such

a division of labor is both �nancial � cheaper to maintain individuals who are ideologically

committed but do not participate full-time � as well as tactical � full-time �ghters tend

to be more skilled and therefore protected from unnecessary risks that would undermine the

insurgent e�ort. Figure 1 shows the estimated number of full and part-time �ghters for a

number of modern insurgencies.

Figure 1

��

�������

��������

��������

��������

�������� ���������� ���������� ���������

�����

�����������������

��������������� ����������������������

�������������

������������������

2The presence of part-time �ghters already poses international law conundrums as to whether these�ghters are protected by the Geneva conventions or not (CITE).

3Source: http://www.globalsecurity.org/military/ops/iraq_insurgency.htm

4

As noted, part-time �ghters greatly outnumber those considered �full-time� or completely

devoted to the insurgent e�ort. In fact, security-based NGOs have recognized the vulnerabil-

ity of these part-time �ghters to economic conditions and have launched initiatives to target

part-time �ghters for �reintegration� (New Strategic Security Initiative 2010, Afghanistan).

Even historically, during the American Civil War (1861-1865) desertions from the Confed-

erate Army increased in the months of June and July, the harvesting times for tobacco � an

important Southern crop at the time (Giu�re 1997). In addition, during the Russian Civil

War (1917-1922) desertion rates in the Red and White armies � largely formed by peasants

� were notoriously high during the summer harvest (Figes 1996 cited by Dal bo and Dal

bo 2011: 657). This is consistent with the ubiquitous presence of opportunity costs e�ects

across con�ict settings.

Contribution. The paper contributes to the literature in several ways. First, the paper

shows that labor availability shapes the violent activities of insurgent groups, even those that

are highly ideological such as the Taliban or the Iraqi insurgency. Given these insurgencies

heavily rely on part-time �ghters � who shift between con�ict and legal work � their labor

availability is in�uenced by these changing economic opportunities. The robustness of these

results across di�erent contexts and datasets is important given the well-known mixed results

in the con�ict literature.

Second, the paper contributes to the on-going debate by providing a rationale for why

we might observe mixed results in the literature: these may actually be driven by the high

persistence of the shocks analyzed. As such, the paper makes a broader methodological point

about how the use of certain income shocks may lead to systematic biases making it di�cult

to capture the mechanism of interest and lead to seemingly contradictory �ndings.

Finally, the paper provides a novel source of exogenous variation in the demand for labor

to study the mechanisms a�ecting con�ict. Because harvest increases the static opportunity

costs of �ghting, while keeping the dynamic bene�ts of �ghting constant, it reduces potential

omitted variable bias due to consumption patterns or the perceived returns to victory (Fearon

2008; Chassang and Padro-i-Miquel 2009).

In terms of policy, care should be taken in interpreting our results for the opportunity cost

mechanism as evidence in favor of employment programs or permanent forms of development

aid. While there may be other reasons why these policies should be in place, their persistence

across periods may lead to unintended consequences. For example, a permanent wage or

5

employment subsidy may mean that households are wealthy enough to devote time to �ght

causes they care about. Or, they may encourage people to �ght to capture the rents from

these schemes. Indeed, recent studies highlight how common it is for insurgent groups to

appropriate aid which in turn leads to greater armed con�ict (Nunn and Qian 2014; Crost

et. al. 2014).4 One possibility is to create policies that are both temporary and anticipated

that would neutralize their impact on con�ict.

2 Theoretical Framework

A large empirical literature typically uses changes in commodity prices as an instrument for

income to assess its e�ect on con�ict. We study an alternative driver of the opportunity

cost of �ghting: the variation in labor demand due to the timing of harvest. In this section

we compare our estimates of the opportunity cost of con�ict to those in the literature to

examine whether any one measure is better at uncovering the �true� e�ect.

Speci�cally, we compare estimates of opportunity cost parameters from persistent shocks

versus seasonal shocks to labor supply in some simple models of con�ict inspired by the

main motivations for violence in the literature. For each model, we provide a precise def-

inition of the �true opportunity cost� of violence � mainly, the elasticity of time spent on

con�ict activities with respect to wages keeping everything else constant � and compare it

to estimates from a regression of violence on wages using model-generated data driven by

persistent (e.g. commodity prices) versus seasonal variation in wages. In our main model

(Section 2.1), rebels engage in violence in order to capture a resource which has some mon-

etary value (a �greed� model). In Section 2.2, we sketch a model where rebels engage in

violence for a cause (a �grievance� model), but leave the details of the dynamic model to the

appendix. The intuition of the �greed� model is also easily extended to a situation where

households provide counterinsurgency information in exchange for payment (Section 2.3).

It turns out that standard regressions with persistent (non-seasonal) unanticipated shocks

lead to upward biased estimates of the opportunity cost of violence, because other factors

determining violence also covary with wages. These other factors are model speci�c: in our

main �greed� model, bene�cial economic shocks that increase wages also increase value of

spoils of wars, which (by itself) tends to increase the time allocated to violence (Fearon 2008;

Chassang and Padro-i-Miquel 2009). In the grievance model, higher wages also make the

4Other recent examples of aid-theft cited by Nunn and Qian (2014) are Afghanistan, Ethiopia, SierraLeone, among others.

6

agent wealthier, which reduces the marginal utility of consumption, increasing the relative

value of an extra unit of time allocated to violence.5 An increase in persistence exacerbates

this bias. Because commonly used commodity prices in the literature are highly persistent

they will tend to underestimate the role of opportunity costs mechanisms in con�ict, which

helps to rationalize the wide variety of estimates in the literature.6

In contrast, seasonal shocks are both temporary and anticipated, which means that other

factors determining violence tend to be held constant, even though opportunity costs change.

This creates an almost-ideal environment in which even simple regressions without controls

can isolate the true e�ects of changes in the opportunity costs of violence. The reason

other factors are held constant is because they are forward-looking variables. For example,

the value of an asset captured by rebel armies in a �greed� model depends on the present

discounted value of future cash �ows generated by the asset (e.g. oil). Anticipated changes

in earnings do not a�ect asset prices. In the grievance model, consumption doesn't respond

to anticipated or temporary shocks, keeping the marginal utility of consumption constant.

For these results, shocks only need to be temporary or anticipated � but seasonal shocks

are both.

2.1 Greed Model

One of the most popular motivations for con�ict in the literature is a contest for resources

(Haavelmo 1954; Hirshleifer 1988, 1989; Gar�nkel 1990; Skarpedas 1992; Gar�nkel and

Skaperdas 2007). In this section, we present the one side of a �contest� model, where rebels

are �ghting for control of economic pro�ts and the probability that they win is increasing in

their e�ort devoted to �ghting. For tractability we keep constant the strength of counterin-

surgency forces. In our model, e�ort is the time that seasonal �ghters devote to con�ict,

which they could otherwise devote to working at wage W . The seasonal �ghter balances

the extra income they could get working against the greater chance they will win if they

5In the counterinsurgency information model, higher consumption from persistent shocks lowers themarginal utility of consumption which reduces their willingness to provide tips, potentially increasing insur-gent violence.

6We thank Scott Ashworth for this observation. The quarterly persistence of oil and co�ee prices isaround 0.96. Speci�cally, this is a regression lpriceRt = ρlpriceRt−1 + Ξt over 1960Q1-2015Q2 for averageoil prices (ρ = 0.97), Arabica Co�ee (ρ = 0.95) and Robusta Co�ee (ρ = 0.97) taken from World Bank PinkSheet., with nominal prices de�ated by the US CPI (data from FRED). Results do vary over sub-samples,but commodity prices are still highly persistent. For example over 1988-2005, ρ = 0.9 − 0.94 for thesesame shocks. Rainfall shocks are unsurprisingly not very persistent, though often have limited e�ect onagricultural output due to presence of irrigation.

7

spend that time �ghting. If economic pro�ts are constant, then an increase in wages makes

working relatively more attractive and �ghting less attractive. However, as pointed out by

Fearon (2008) and Chassang and Padro-i-Miquel (2009), the same shocks (e.g. productivity

shocks or commodity price shocks) can increase both the costs (foregone wages) and ben-

e�ts (pro�ts) of �ghting, and so have no net e�ect on violence. In a dynamic setting, the

costs of �ghting are incurred today, whereas the bene�ts of winning are potentially in the

future, such that negative temporary shocks increase violence more than persistent shocks

(Chassang and Padro-i-Miquel 2009). As such, seasonal labor demand allows for a clean

identi�cation of the true opportunity cost of violence, because seasonal variation in wages

are temporary and predictable, meaning that the potential spoils of winning are constant in

high versus low labor demand seasons.

Related literature Our model relates to Fearon (2008), Chassang and Padro-i-Miquel

(2009) and Dal bo and Dal bo (2011). In Fearon's (2008) baseline model, there are no dy-

namics, and the rebels choose the optimal size of their forces, given the marginal cost of

recruitment and the government's response function. Con�ict is unavoidable and a larger

force increases the probability of winning, which then allows the winner to tax at a given

rate.7 Chassang and Padro-i-Miquel (2009) present a bargaining model where two players

decide to {attack, not attack} rather than choosing the intensity of con�ict, conditional on

a �xed labor cost of �ghting, and an o�ensive advantage. If the rebels win, they gain the

resources of the other side and in the dynamic version, winning is decisive forever. Dal bo

and Dal bo (2011) presents a two-industry, two-factor static trade model with an appropria-

tion sector to show how sector-speci�c prices a�ect con�ict. Our model includes ingredients

from all of these models. Like in Fearon (2008), con�ict varies at the intensive rather than

extensive margin. Like Chassang and Padro-i-Miquel (2009), the gains from winning are

dynamic whereas the costs are static, meaning that temporary but not permanent produc-

tivity shocks a�ect violence (winning is also decisive). Like Dal bo and Dal bo (2011), our

appropriation/�ghting technology is strictly concave in labor (re�ecting congestion e�ects);

our production function is non-linear in labor such that real wages depend on the allocation

of labor; and we abstract from the government's response to violence.

7In later models, Fearon (2008) add a detection probability, di�erent abilities of rebels and governmentsto tax, and changes the contest function to a �capture� function.

8

2.1.1 Static Model

The household has one unit of time and decides at the start of the period how to split it

between working or �ghting. If the rebels win the �ght, the agent earns the economic pro�ts

from production, Π. These pro�ts can be thought of as the returns to a �xed factor like

land, capital or a natural resource. If the rebels lose, the part-time �ghter gains nothing.

Whether the rebels win or lose, the agent still collects labor income from working (1−V )W .

The probability that the rebel win is increasing but concave in the time allocated to violence

V :

p = ψV 1−γ (1)

0 < γ < 1 governs the e�ectiveness of the �ghting technology, which means that the

p′(v) = ψ(1−γ)V −γ is decreasing in V .8 A nice feature of this function is that the �rst hour

of time devoted to con�ict is in�nitely productive (i.e. limv→0p′(v) = ∞), which captures

the stylized fact that many countries have a low-level insurgency with very little chance of

overthrowing the government (Fearon 2008).

The household's problem is:

maxV pU(cwin) + (1− p)U(closes)

such that

cwin = W (1− V ) + Π

close = W (1− V )

Output is produced using only labor (1− V ), and labor is paid its marginal product W .

As labor markets are competitive, the household takes the wages and pro�ts as given. A

is total factor productivity, which is the key exogenous variable in the model. If household

produced a cash crop for export, and consumed only imported goods, then A = pY /pC could

capture the terms of trade used when output, wages and pro�ts (Equations 2-4) are written

in terms of the consumption good.

Y = A(1− V )α (2)

8The strength of the counterinsurgency is governed by ψ. Restricting 0 < γ < 1 also keeps the objectivefunction concave.

9

W = αA(1− V )α−1 (3)

Π = Y −W (1− V ) = (1− α)A(1− V )α (4)

To separate the mechanism from the one in a the grievance model (below and in the

appendix), we make three assumptions (i) U(C) = C (linear utility, risk neutral agents), (ii)

No saving or borrowing, (iii) Violence is NOT in the utility function.

Substituting for p and U(C), the HH's problem becomes:

maxVW (1− V ) + ψV 1−γ︸ ︷︷ ︸ProbWin

Π

The FOC is:

V =

[ψ(1− γ)

Π

W

] 1γ

(5)

Taking logs, we can get an equation to take to the data (actual or simulated):

lnV =1

γlnψ(1− γ) +

1

γlnΠ− 1

γlnW (6)

De�nition. The opportunity cost of violence is the elasticity of violence with respect to

wages, keeping everything else constant,∂lnV

∂lnW= −1

γ.

In order to estimate −γ−1 from a regression of violence on wages in Equation 6 requires

controlling for lnΠ. If instead researchers ran a univariate speci�cation, lnΠ would be

subsumed into the error term. As lnW and lnΠ tend to be positively correlated (see below),

this will bias upwards the coe�cient on wages.

To see this, suppose that changes in wages are driven by changes in productivity A (or

alternatively, the terms of trade). By substituting in lnW and lnΠ, Equation 5 becomes

Equation 7. One can see that an increase in productivity (A) increases Π and W propor-

tionately, and so in Equation 7 A cancels out exactly and violence is constant (i.e. A does

not appear in Equation 7). This mean that if one ran a regression of lnV on lnW , one would

get a coe�cient of zero, rather than −γ−1. This is Fearon's (2008) result that economic

development increases both the opportunity cost of violence as well as the spoils of war,

leaving the level of violence unchanged.

10

V =

[ψ(1− γ)

(1− α)(1− V )α

α(1− V )α−1

] 1γ

(7)

2.1.2 Dynamic Model

Seasonal variation in productivity provides a context where the opportunity cost of violence

changes, but the value of the prize of �ghting is approximately constant. This e�ectively

removes the omitted variable bias described above, allowing an unbiased estimation of the

�true opportunity cost� parameter −γ−1 even if we can't observe Π. The opportunity cost

of �ghting varies with seasonal changes in productivity because it is incurred contempora-

neously. In contrast, in a dynamic setting the bulk of the �prize of winning� are future rents

from resources captured, which will be almost constant across �seasons� because changes in

productivity are temporary and anticipated. In contrast, persistent shocks like commodity

prices raise both the prize and cost of �ghting, leading to upwards biased estimates of the

opportunity cost of �ghting.

More formally, let VL(A) be the value (discounted lifetime expected utility) of a part time

rebel �ghter not in power deciding how much time to devote to �ghting versus working. The

state of the economy is A (total factor productivity) and whether the rebels are in power (L

for lose summarizes their past defeats). If the rebels win, they will gain pro�ts today Π and

the value of being in power next period VW . This value depends on next period's productivity

A′ (next period is denoted with ′). Like Chassang and Padro-i-Miquel (2009), we make the

simplifying assumption that if rebels win they stay in power forever.9 If the rebels lose,

tomorrow the part-time rebel faces the same problem, and so have the same value VL(A′).

The probability of winning, as in the static model, is p = ψV 1−γ. β is the quarterly discount

rate. The household has linear utility in consumption, cannot save/borrow, and does not

intrinsically value violence. W (1 − V ) is the income received from working (regardless of

whether the rebels win or lose).

VL(A) = maxVW (1− V ) + ψV 1−γ︸ ︷︷ ︸ProbWin

(Π + βE [VW (A′)]) + (1− ψV 1−γ︸ ︷︷ ︸ProbLose

)βE[VL(A

′′)]

If the rebels win, then there is no gain from �ghting anymore, and so seasonal �ghters

9An alternative version includes an exogenous loss of rebel control with probability 1− δ. For low valuesof 1− δ, the model produced similar results (for high values it sometimes did not solve). But this makes themodel much more complicated.

11

spend all their time working (V = 0). As before, they earn labor income W (1 − V ) = αY

and also control pro�ts Π = (1 − α)Y , yielding total income Y . However, we also allow

for rebel controlled production to be less productive by a factor 0 < λ ≤ 1 such that

Y = λA(1− V )α = λA. As such, the value of a part-time �ghter when they are in power is:

VW (A) = λA+ βEVW (A′)

The exogenous process for productivity is given by Equation 8 if there are persistent

productivity or commodity price shocks, or Equation 9 when there is seasonal variation in

productivity.

For persistent shocks:

lnA′ = ρlnA+ e (8)

Or, for seasonal shocks:

lnAL = lnA for t+ 1, t+ 3, ... (9)

lnAH = lnA+ χ for t, t+ 2, t+ 4, ...

The �rst order condition is:

W = (1− γ)ψV −γ [Π + β [EVW (A′)− EVL(A′)]] (10)

On the left hand side is the gain from devoting an extra hour to working: wages. On

the right hand side is the gain from an extra unit of violence: the change in the probability

of winning p′(V ) = (1 − γ)ψV 1−γ times the prize of winning: pro�ts today Π, and the

discounted di�erence in future utility from being in power VW (A′) relative to not being in

power VL(A′).

Model solution and simulation Log-linearizing the model around the non-stochastic

steady state (where A′ = A = A), the losing value function, FOC, and winning value

function become Equations 11, 12 and 13 respectively.10 Here a lower case variable with a

hat (x) represents the percentage deviation from steady state (which are denoted in capitals

10This is a �rst order Taylor series approximation of the model's FOCs and value functions. The �log�part refers to the fact that we perform the Taylor's series approximation with respect to logXt rather thanXt (i.e. rewrite Xt = elogXt)).

12

X). If one could control for the prize of winning (ΠΠ +β(VWEv′W − VLEv′L)), one could run

a regression of violence (v) on wages (w) which Equation 12 suggests one would estimate the

true opportunity cost parameter −γ−1. But as the value of the prize of winning is typicallyunobserved and is correlated with wages, we use the model to calculate the degree of omitted

variable bias for di�erent types of shocks.

VLvL = wW (1−V )−vW V +(1−γ)vψV 1−γ [Π + βVW]+ψV 1−γ [Ππ + βVWEv

′W

]+βVLEv

′L (11)

v = −1

γw +

1

γ

ΠΠ + β(VWEv′W − VLEv′L)

Π + β(VW − VL)(12)

vW =A

VWa+ βEv′W (13)

where the marginal product of labor and the value of pro�ts are:

w = a+ (1− α)V

1− Vv (14)

π = a− α V

1− Vv (15)

The model is not analytically tractable, so instead we simulate data when productivity is

driven by persistent shocks (like commodity price shocks) or anticipated temporary seasonal

variation in productivity, and estimate a regression of simulated violence on simulated wages.

We calibrate −γ = −13to match the estimated elasticity of violence with respect to wages

found in Colombia (-1.5).11 Speci�cally, we use an indirect inference approach and choose γ so

that our estimated coe�cient on simulated data with shock persistence ρcoffee = 0.96 (similar

to the quarterly persistence of co�ee prices) matches -1.5, which is what we empirically �nd

with the available data.12 As before, with a persistent shock of ρ = 0.96, the bias due to

11Yearly log wages are instrumented by co�ee prices x co�ee suitability. Estimated at the municipal levelwith �xed e�ects. We drop zero violence municipalities. Data are from Dube and Vargas (2013), thoughthis is our own regression, not the ones that the authors estimate (the authors use wages as a dependentvariable).

12Other parameters: α = 0.5 is calibrated to the all-countries, all-years average of the labor share fromPWT8 (full value 0.5459). λ = 3/4 and ψ = 0.015 are chosen to keep the steady state share of violence low(at around 7%), while matching the elasticity of violence to wages in the data with ρ = 0.96. β = 0.99implies an annual real interest rate of around 4%. Steady state values of A = 1 and V = 0.07. As discussed

13

omitting variation in the value of the prize of winning is substantial: the estimated coe�cient

of -1.5 is around half of the true value of −3 = −γ−1 (Figure 2 LHS, blue line). The bias

is small for very transient shocks, but rises sharply as shocks become persistent. In fact,

as shocks become perfectly persistent, the estimated elasticity of violence with respect to

wages becomes positive. In contrast, a regression of violence on seasonal variation in wages

almost exactly uncovers the true opportunity cost parameter (-2.98 (green line) versus a

true value of -3 (red line)). Examples of simulated paths of violence, wages and the prize of

�ghting are shown on the RHS of Figure 2: in the persistent shocks simulation (top panel),

the prize of �ghting rises slowly in the middle of the simulation and then falls.13 In the

bottom panel, the prize of winning is almost completely una�ected by seasonal movements

in productivity, which is what allows us to uncover the true opportunity cost parameter with

a simple regression of violence on wages.

0 0.2 0.4 0.6 0.8 1−3.5

−3

−2.5

−2

−1.5

−1

−0.5

0

0.5

1

Persistence of Wage Shock (Quarterly)

Ela

stic

ity o

f Vio

lenc

e to

1%

incr

ease

in W

age

True Opportunity Cost (−1/γ)=−3Regression Coeff. (AR(1) Simulations)Regression Coeff. (Seasonal) [= True Opp Cost]Persistence Commodity Prices in Data (Vertical Line)

5 10 15 20 25 30 35 40−0.2

−0.1

0

0.1

0.2Persistent shocks (ρ=0.96); "Greed" Model−Generated Data

5 10 15 20 25 30 35 40−0.2

−0.1

0

0.1

0.2Seasonal Labor Demand; "Greed" Model Generated Data

Violence (deviation from SS)Wages (deviation from SS)(1/γ)Prize from fighting (deviation from SS)

Figure 2: Greed Model: Panel A: estimated coe�cient (LHS) and Panel B: simulateddata (RHS)

2.2 Grievance model

In this model, we assume that rebels engage in violence for some �grievance� in which they

place intrinsic value: examples include ethnic or religious hatred, retaliation, or nationalism

above, γ = 1/3 implies a true opportunity cost of -313The prize of the �ghting moves slowly because it is forward looking: in steady state pro�ts today are

only around 2% over the value of winning (Π/(Π + β(VW − VL)).

14

(Horowitz 1985). That is, rebel violence is in the utility function. To make the mechanism

completely clear � and to di�erentiate it from the �greed model� � we assume that there

are no monetary bene�ts from violence, and to keep the model tractable we do not model

the government's response. A key assumption is that households get diminishing marginal

utility from allocating additional time to violence (UV > 0; UV V < 0), which means that an

increase in �opportunity cost� will lead to a reduction in time allocated to violence, other

things equal.14 We sketch a static model here, and reserve the dynamic model � which

introduces seasonality and persistent shocks � for the appendix.

Static Model

As before, consider the problem of a household who has an endowment of one unit of time

to divide between �ghting V and working (1− V ) at an exogenous wage W . More formally:

maxV,CU(C, V ) such that C = W (1− V ) (16)

Assuming an interior solution, the household's �rst order condition is:

UV = UCW (17)

Equation 17 says that the marginal utility from spending an extra hour �ghting (LHS),

must be equal to the hourly wage weighted multiplied by the contribution of consumption

to utility (RHS). An increase in wages by itself means UV must increase, which implies lower

violence as UV V < 0 (the �substitution e�ect� or opportunity cost channel). However, an

increase in wages will also usually increase consumption and reduce UC (the marginal value

of extra income in terms of utility, UCC < 0), such that UV falls and violence increases

(the �income e�ect�). Which e�ect dominates depends on the parameters of the model,

but so long as income e�ects are positive, violence will move by less than the opportunity

cost/substitution e�ect suggests.

Assuming a standard constant relative risk aversion utility function , U(C, V ) = C1−σ/(1−σ) + ψV 1−γ/(1 − γ) with σ ≥ 0, 0 < γ < 1 substituting and taking logs, we get a simi-

lar expression for violence as Equation 6 in the Greed model. As before, the opportunity

cost is the elasticity of violence with respect to wages, keeping everything else constant, or

14 We also assume that violence and consumption are separable, that is UCV = UV C = 0. This lastassumption means that the marginal utility of �ghting does not depend on how rich one is. Concavity alsoallows us to assume limV→0UV =∞, which corroborates the prevalence of low-level insurgencies describedin Fearon (2008) �- even if the cause is not so convincing.

15

∂lnV/∂lnW = −γ−1.

lnV = −1

γlnψ +

σ

γlnC − 1

γlnW (18)

Despite the di�erent motivation for violence, the grievance model has a very similar

omitted variable problem as the greed model above. The analogue of the unobserved value

of the prize (Π) is consumption C (which determines the marginal utility of consumption) in

Equation 18 and is usually not observed (or is poorly measured). As such researchers might

be forced to estimate some variety of lnV = β0 + β1lnW + et, where et = σγlnC. However,

typically cov(lnC, lnW ) = σCW > 0 � people on higher wages have higher consumption

and a lower marginal utility of consumption � and so the estimated magnitude of the

opportunity cost of violence will be upward biased (towards zero).15 In the special case that

utility is linear (σ = 0), all income e�ects are removed and a simple uni-variate regression

of violence on wages uncovers the true opportunity cost −1/γ, regardless of movements in

consumption.

Seasonal vs persistent shocks and the permanent income hypothesis (PIH)

In the dynamic model in the appendix, consumption is determined by the permanent income

hypothesis � agents smooth their consumption over time and only consume out of their

permanent income.16 This means that anticipated or temporary shocks to income/wages

will be smoothed by savings/borrowing and will have almost no e�ect on consumption.

As seasonal shocks are both anticipated and temporary, σγlnC in Equation 18 will be kept

constant, yielding an unbiased estimate of the opportunity cost mechanism from a regression

of violence on seasonal wage shocks � even if lnC is unobserved.

In contrast, highly persistent increases in wages will lead to a large increase in permanent

income, which will increase consumption. With log preferences (σ → 1), a permanent shock

will raise consumption in proportion to wages, which will mean permanent labor demand

shocks have no e�ect on violence, leading to a estimated opportunity cost of zero if the

researcher can not control for consumption.17

15That is, Eβ1 − [−1/γ] =σ

γ

σWC

σ2W

> 0

16That is log-linearized Euler equation implies ct ≈ Etct+1.17The larger is σ, the stronger are income e�ects and the larger the bias for permanent wage shocks.

For permanent shocks, one can take a log-linear approximation of of FOC and budget constraint to yield:

v =(σ − 1)

γ + σV /(1− V )w. If 0 < σ < 1, the substitution e�ect dominates the income e�ect: the coe�cient on

wages is still negative, but is biased upwards. However if σ > 1 � such as σ = 2 for numerical simulations in

16

In the appendix, we generate simulated data in the grievance model with seasonal and

persistent commodity shocks, and run a univariate regression of simulated violence on simu-

lated wages. As in the dynamic greed model above, regressions on seasonal shocks are able

to uncover the true opportunity cost, but regression on data driven by commodity prices

shocks are substantially upward biased because commodity prices are highly persistent.

2.3 Counter-insurgency and the value of information

Berman et al (2011a) argue that information is a key component of any counterinsurgency

strategy: if government forces do not receive information on where the rebels are hiding

(for example), then counterinsurgency e�orts will be ine�ective. In other words, military

e�ort and information are complements. In order to gain information, government forces

often pay locals for tips. Berman et al (2011b) argue that this provides a reason why

they �nd a negative relation between unemployment and violence: when unemployment

is high, it is cheaper for government forces to buy information from the local population,

which then reduces insurgent violence. The fact that the local population does not provide

information freely suggests that there is some sort of utility cost to providing it (e.g. they

don't like �snitching�, or it is dangerous). Hence, the willingness of the household to provide

information depends on its marginal utility of consumption, which could fall with positive

persistent shocks, but is kept constant by seasonal shocks. We brie�y sketch the argument

below, as it is almost identical to the mechanism in the grievance model above.

Consider a modi�cation of the static set up in Equation 16 above to incorporate in-

formation provided to counter-insurgency forces I. The household doesn't like to provide

information, so UI < 0, and dislikes each additional unit of �snitching� even more, such that

UII < 0 (we continue to assume that utility is separable in information, consumption and

time allocated to violence). The household gets a payment s for each �snitch�, which we

assume is constant. The household's problem is then:

maxV,IU(C, V, I) such that C = W (1− V ) + sI (19)

The FOC wrt to time allocated to violence is unchanged from Equation 17 above, whereas

the appendix � a permanent increase in wages reduces the marginal utility of consumption su�ciently thatan increase in wages actually increases violence. However, commodity price shocks are highly persistent butnot permanent, and as such simulations suggest that an increase in wages due to a persistent commodityprice shock still reduces violence, though by half as much as the true opportunity cost mechanism wouldsuggest.

17

the FOC wrt I implies:

− UI = UCs (20)

One can see that if there is an increase in consumption from a persistent shock (such as

a persistent commodity price shock), then UC will fall (because UCC < 0). As s is constant

−UI also must fall. Note that −UI > 0 and −UII > 0, so the only way for −UI to fall is

for the household to provide less information: richer households have less need to become

an informant as in Berman et al (2011b), which could actually increase aggregate violence.

But because seasonal shocks are temporary and anticipated they will not lead to a change

in consumption, and so UC will be constant, and information provision will be una�ected.

As before, this allows seasonal variation in wages to produce an cleaner estimate of the

opportunity cost of mechanism.

3 Data and Empirical Methodology

The results described above lead to the following empirical implication: the onset of harvest

has a negative impact on con�ict intensity by increasing the returns to working (e.g. wages)

relative to �ghting. To bring this implication to the data one would ideally instrument the

variation in monthly wages driven by harvest and examine its e�ect on con�ict. In practice,

con�ict-ridden areas (and even non-con�ict ones) often lack comprehensive monthly time-

series for local wages. Hence we focus on estimating the reduced-form e�ect of violence

on harvest onset. The idea is that a negative coe�cient would be consistent with the idea

that increases in local labor demand reduces the attractiveness of �ghting. We also provide

additional evidence showing that harvest brings about changes in local labor markets to

support the idea that the e�ect is driven through this mechanism.

3.1 Data

The data for our con�ict episodes relies on a number of di�erent sources. For every con�ict

episode we sought disaggregated data on violent incidents to match the spatial variation

of harvesting calendars across the country. Because we exploit monthly-by-district changes

in labor markets and include a number �xed e�ects indicators, the only factors that could

confound the e�ect observed are those which vary at the district-by-month level (for example,

precipitation or temperature).

18

Violence. Our main dependent variable is the monthly (m) share of attacks per district

(i) (AttacksimtAttacksit

) relative to its total given a year (t). This is a way to normalize across di�erent

con�ict settings. We look at di�erent con�ict settings and datsets on violence, as a way

to avoid assigning disproportionate weight to a single data collection procedure given the

well-known di�culties in recording violence. We use both very precisely geolocated datasets

(e.g. latitude, longitude) as well as those in which the level of aggregation is that of small

administrative units (e.g. districts or municipalities). However, we generally rely on the

district level results as a way to reduce measurement error. For Iraq we use the World

Incident Tracking System (WITS), the Global Terrorism Dataset (GTD) and Iraq Body

Count dataset (IBC). In the case of Pakistan, we use the BFRS dataset on political violence

which is available at the district-level as well as the GTD data which is precisely geolocated.

In the case of Afghanistan, we rely on con�ict data provided by WITS between 2004 and

2010 aggregated at the level of the district.

Harvesting Calendars. For the case of Afghanistan, Pakistan and Iraq, the timing of

harvest for each cell or district is provided by the FAO Global Agro Ecological Zones v3.018

(GAEZ v.3.0) which provides high resolution maps for the start and length of the growing

cycle for a number of crops. Our harvesting indicator takes the value of 1 for the month

immediately after the end of the growing cycle. For each crop we also capture whether it

utilizes high, medium or low inputs which indicates whether the crop is rain-fedo or irrigated.

Because our indicator captures the onset of harvesting for any type, districts could have more

than one harvesting month if it cultivates more than one type of crop and these di�er in

their harvesting date.

In the case of planting, we follow the same approach and create an indicator for the

month prior to the start of the growing season under the logic that this is the time in which

land is prepared and sowed before seeds can grow. As an example, Figure 3 below shows

the harvesting calendars for Iraq. Since the harvesting month varies across districts within

the year, it provides within country variation in the month in which wheat is cultivated thus

allowing for identi�cation of its e�ect. For Iraq, around half the wheat is cultivated in June,

yet, some areas also harvest as late as September and others as early as April.

18Available at: http://www.gaez.iiasa.ac.at/

19

Figure 3: Harvesting Calendar Iraq

Crop Intensity. Crop intensity is measured in hundreds of square kilometers and is

calculated by the FAO for the perid 1960-1990, which clearly precedes our period under

study. We interact the harvesting and planting indicator with the historical intensity of crop

production to avoid giving greater weight to areas with little to no crop production.

To illustrate, Figures 4 through 6 show the raw images provided by GAEZ v.3.0 and

those once linked to a 0.1 by 0.1 decimal degrees grid for the Iraqi case (approximately

11kms by 11kms cells). Figure 4 shows the intensity of wheat production; Figure 5 shows

the start day cycle for medium input crops and Figure 6 shows the length of the cycle. The

weighted harvesting calendar is thus determined by when wheat is planted combined by how

long it needs to grow according to where it is cultivated and weighted by how much wheat is

cultivated. This �ne-grained information is then aggregated at the district-level to calculate

the intensity with which a given district is �in harvest�.

Figure 4: Left: Wheat Production. Right: Grided Production

20

Figure 5: Left: Start Day Medium Input Wheat Irrigated. Right: Gridded Start Day

Figure 6: Left: Length of Cycle Medium Input Wheat Irrigated. Right: Gridded LengthCycle

Additional controls. Additional control variables at the cell or district level always

include those of precipitation and temperature. Although the timing of harvesting is unlikely

to be in�uenced by crop production, it is possible that monthly factors determining harvest

may also a�ect the intensity of violence thus confounding our results. Therefore, we collected

data on monthly-district measures of precipitation (in millimeters) and temperature (degrees

Celsius) for Iraq, Afghanistan, and Pakistan provided by Willmott and Matsuura (2001).

To examine the e�ect of harvesting on local labor markets we also examine household

surveys which ask for monthly patterns of employment and time use, which are designed

to be representative of the rural sector. While the survey asks for monthly employment

21

patterns, unfortunately it does not do the same for wages. In the case of Iraq, we use the

Living Standards and Measurement Study collected by the World Bank in 2006-2007.19

3.2 Estimation

Our outcome of interest %Attacksimt, is the share of attacks in a district i, calendar month

m and year t relative to the total in that district and year. Our key independent variable

Harvim × Prodi is the number of hundred square kilometer of wheat in harvest in district

i , month m and year t, unless otherwise speci�ed. In all speci�cations we also include the

e�ect of the planting season on con�ict. Hence we estimate:

Attacksimt = αit + γm + β(Harvim × Prodi) + ximt + eimt (21)

Where αit is a district by year �xed e�ect, and γm is a month �xed e�ect (e.g June);

ximt is a vector of monthly district characteristics such as monthly temperature in degrees

Celsius and precipitation in millimeters. The parameter of interest is β which captures the

e�ect of harvesting on con�ict intensity. Standard errors are clustered at the district level,

which accounts for serial correlation in the error terms for that spatial unit.

3.3 Threats to Identi�cation

Our identi�cation strategy exploits the fact that seasonality or the timing of harvest is clearly

exogenous to the intensity of armed con�ict. That is, we exploit the roll-out of harvest and

compare how violence changes in districts in harvest relative to months without it. Since the

timing of harvest is given by a combination of geographic and climate factors, it is unlikely

to be manipulated by con�ict dynamics. Certainly con�ict may a�ect crop production itself,

yet, this would only run against �nding any relationship between the harvesting month and

the intensity of armed con�ict within a district.20

While reverse causality is not necessarily a concern, a more important challenge comes

from omitted variable bias or time-varying determinants of harvest (e.g. precipitation, or

temperature) which may correlate with con�ict. For example, in the Iraqi case Figure 7

below shows how the onset of harvesting (roughly from May to July) is indeed accompanied

19Available at: http://econ.worldbank.org/20For instance, con�ict may shift grain collection for some weeks, yet, it is unlikely to do so for a whole

month (which is our the size of our indicator �window�) as it would be pointless from the producer standpoint:either crops will not be ripe or they would rot as time passes.

22

by an increase in temperature and a decrease in precipitation. If temperature were to have

a positive e�ect on con�ict, as a number of studies suggest (Burke et. al.2009; Hsiang et. al.

2013), this would only exert an upward bias in our results. That is, the true coe�cients would

be actually larger (e.g. more negative) than our estimated coe�cients. Similar concerns arise

with the amount of precipitation, since intense rainfall may constitute a physical impediment

to conducting attacks. However, as shown in the LHS of Figure 7, precipitation is actually

lower at times in which most of the harvesting is occurring such that, if anything, coe�cients

would also be upward biased.

Figure 7: Monthly Precipitation (left) and Temperature (right)Patterns in Iraq

4 Empirical Results

If opportunity costs are an important consideration to participate in con�ict activities, it

must therefore be present in cases where part-time �ghters are common (or where there is

a large share of individuals deciding whether to �ght or not). In this section we show how

across di�erent con�ict settings seasonal labor markets play a key role in determining within-

year variation in the intensity of violence. Given the di�erences in data sources and coding

methods we present each case separately while holding constant the main speci�cation, unless

otherwise speci�ed.

4.1 The Iraqi Con�ict (2004-10)

Between 2004-2011 Iraq was gripped by a civil con�ict along sectarian lines as well as Sunni

insurgencies in numerous parts of the country. The intensity of the con�ict, coupled with

23

the strong reliance on agriculture as an economic activity and the cultivation of wheat as

the main subsistence crop, makes it an ideal setting to explore the importance of seasonal

labor markets for violence intensity.

Iraq Body Count (2004-2009). Our analysis starts by examining the patterns of

insurgent activity using geocoded incidents captured by instances of district-level violence

collected in the Iraqi Body Count dataset (IBC). This dataset is maintained by a non-pro�t

organization which quanti�es the number of casualties based on multiple sources (including

media) and distinguishes between the type of attack such as airstrikes, artillery �re, bomb

devices, gun�re, among others. We use these di�erent categorizations to examine whether

harvest induces insurgent groups to favor certain tactics at the expense of others when labor

availability is low (Bueno de Mesquita 2013). Speci�cally, we distinguish between labor

intensive attacks, or those that require greater manpower to be carried out (e.g. armed attack

or assault), and asymmetric attacks, those in which participants are not able to exchange

�re and have generally lower manpower requirements (e.g. IEDs) (Bueno de Mesquita et.

al. 2015). We also report results where we pool across all attack types.

Columns (1) to (4) of Table 1 above shows that in this dataset there is evidence of

lower seasonal attacks during harvest periods. Speci�cally, an increase of a hundred square

kilometers of wheat production at harvest is associated with a reduction in the intensity of

attacks, particularly those labor intensive (direct �re and selective targets) as opposed to

asymmetric ones (indirect �re and bombing). Speci�cally, column 1 shows that an increase

of a hundred square kilometers of wheat cultivation in the district at harvest leads to a

reduction of 0.83 percentage points in reported events. Given the average wheat cultivation

intensity per district is 1.2 hundred square kilometers, the coe�cient entails a reduction of

12.5% in the average monthly share of lethal events captured by this dataset.

24

Table 1: Seasonal Labor and Violent Incidents in Iraq

(1) (2) (3) (4) (5) (6) (7) (8)

Iraq Body Count WITSDistrict-Level Analysis

DV: Monthly % of... Total Attacks Asymm Laborint Victims Total Asymm Laborint Victims

Harvim × Prodi -0.833*** -0.610 -0.921** -0.816*** -0.601* -0.879** -0.083 -0.633**(0.288) (0.536) (0.362) (0.303) (0.327) (0.416) (0.389) (0.313)

Plantim × Prodi -0.323 -0.635* -0.477 -0.416 0.169 -0.255 0.544 0.174(0.346) (0.344) (0.428) (0.342) (0.339) (0.343) (0.426) (0.345)

Mean Harv Area 1.254 1.306 1.272 1.260 1.257 1.275 1.283 1.257Mean DV 8.333 8.333 8.333 8.333 8.333 8.333 8.333 8.333Avg E�ect -12.53 -9.564 -14.05 -12.34 -9.069 -13.45 -1.272 -9.544Observations 5,148 3,240 4,212 5,112 5,040 4,140 4,464 5,040Clusters 92 65 86 91 88 70 86 88DistXYear FE Y Y Y Y Y Y Y YMonth FE Y Y Y Y Y Y Y YTemp & Precip Y Y Y Y Y Y Y Y

Clustered robust standard errors at the district level in parentheses. Prodci is measured in hundred sq kilometers. DV in percentagepoints. *** p<0.01, ** p<0.05, * p<0.1

Cross-validation: WITS Dataset (2004-2010). As a cross-check to our results we

run the same speci�cation but now using as dependent variable insurgent activity captured

by the Worldwide Incidents Tracking System (WITS) which is based on media accounts

of terrorist events.21 This dataset focuses on incidents that are both �international and

signi�cant� in nature and is used as a reference point for the State Department (Wigle

2010).22 In addition to tracking the number of terrorist events, the dataset also provides

broad categorizations of the type of terrorist attacks � whether it was an armed attack, an

attack using improvised explosive device (IED), a suicide bomb, among others.

Table 1 shows the estimates from Equation 21 using as dependent variable the monthly

share of violent incidents in the district. Columns (5) through (8) show how the onset of

harvest leads to a reduction in total levels of violence, as well as a reduction in asymmetric

21Available at: http://www.nctc.gov/site/other/wits.html22�International meant any acts that involved the citizens or territory of more than one country. (...)

What constituted a signi�cant act was even fuzzier and was legally left to the opinion of the Secretary ofState,[5] although there were some prescribed rules promulgated by the State Department. For example, asigni�cant attack meant an act of terrorism that either killed or seriously injured a person, or caused USD$10,000 in property damage.� (Wigle 2010)

25

but not labor-intensive attacks. In terms of magnitude, the coe�cient of -0.6 in column (5)

suggests that an increase of one hundred square kilometers of wheat production at harvest

leads to a reduction in the share of monthly attacks of approximately 0.6 percentage points.

Considering the average production of wheat at the district level is 1.2 hundred square

kilometers, the coe�cient implies a reduction of the average monthly share of attacks of

10%. Similar e�ects are shown in column (6), where the coe�cient of 0.88 also represents

around a 13.5% reduction in the average monthly share of attacks, while column (7) shows

the reduction in labor intensive is very small but not statistically di�erent from zero. These

results closely follow the estimates from the IBC dataset.

Robustness. Additional results in the online appendix23 show that these �ndings are

similar when restricting the sample to only wheat producing areas (Table 2 and 3). It is

worth noting that the onset of planting is either associated with a reduction in attacks or with

a very small coe�cient, but estimates are often less precisely estimated. This lower e�ect is

likely driven by the lower demand for labor posed by planting as opposed to harvesting. In

addition, to make sure the results are not merely driven by the functional form examined,

Table 4 and 5 of the online appendix present the results of regressing an indicator of above the

median wheat production and the harvest calendar (below median production is the omitted

category). As shown, coe�cients are much larger but less precisely estimated for the IBC

dataset. In addition, tables 6 and 7 of the online appendix includes lags for harvesting

an shows that the negative e�ect is driven by the contemporaneous change in harvesting

status, particularly for the IBC evidence. Thus providing little evidence of anticipation

e�ects by armed groups. Finally, tables 8 and 9 of the online appendix compares the e�ect

on violent of harvest in rain-fed versus irrigation areas and shows little di�erences on their

e�ect on con�ict. If anything, the IBC dataset suggests that in Iraq, the variation in irrigated

harvesting areas is driving the e�ect observed in con�ict.

In addition to WITS we also use the Global Terrorism Dataset (GTD) for Iraq as a

�nal cross-check of the results obtained. This dataset is maintained by the National Con-

sortium for the Study of Terrorism and Responses to Terrorism (START) at the University

of Maryland and is also based on media reports, yet, exhibits a much lower frequency of

attacks overall. Estimates of Equation 21 using this dataset shows that the coe�cients vary

in sign and are not statistically signi�cant.24 A more detailed investigation of the di�erences

between the GTD and the WITS or IBC data are an area for future research.

23Available at https://sites.google.com/site/jennyguardado/24Results available upon request.

26

Mechanisms and Alternative Explanations

For these results to be consistent with the theoretical framework, it must be that the onset

of harvest leads to tangible di�erences in labor market outcomes. To assess whether this is

the case we use the 2006 LSMS Iraqi household survey, to examine whether regional patterns

of harvesting relate to employment among agricultural workers. Ideally, we would like to

match each respondent to a particular district and follow it throughout the years. However,

due to privacy concerns, the survey only provides a cross-sectional snapshot at the time of

harvest of individual employment at the governorate level in Iraq (of which there are 18),

therefore, this evidence should be taken as indicative of seasonal patterns of employment

until more �ne-grained information becomes available.

Figure 8 shows the di�erence in the probability of employment among rural agricultural

workers (relative to non-agricultural ones) by month. As shown, these di�erences, controlling

for a number of factors, closely follows the harvesting calendar in rural Iraq. This is consistent

with the idea that harvesting a�ects con�ict by in�uencing local labor markets.

Figure 8: Monthly Employment Patterns

Y-axis: coe�cients from a regression of monthly indicators on employment (�Did you work inthis job in month...�?). Additional controls include: individual's age, level of education, gender,household size and language (Arab or not). We include governorate �xed e�ects and clusterthe standard errors at the level of the survey cluster.

Job Switching and Migration. Although employment patterns mirror the harvesting

calendar in Iraq, it is important to rule out the possibility that individuals switched jobs

within the year. Of the 11,157 individuals surveyed living in rural areas only 521 individuals

or 4.67% reported more than one occupation throughout the year and 0.23% reported the

27

maximum of three occupations during the year. This shows it is unlikely they will be

switching occupations throughout the year. A related concern is whether individuals migrate

to other areas for work, potentially explaining the observed patterns of con�ict. However,

among agricultural workers, the share of individuals reporting an absence from home for an

extended period is only 3.67%.

Labor availability versus Harvest Income. A di�erent concern with our measure

is whether the time of harvest is instead proxying for the income received as opposed to

actual labor availability. This is unlikely in the Iraqi context because most farmers sell their

grain to the governmental Iraqi Grain Board who subsidizes wheat production. Once a year

farmers take their harvest to one of the numerous silos across the country. This takes place

once all harvest is collected due to logistic and transportation costs. Farmers then receive a

receipt which has to be cashed in a bank.25 The process ensures that the time of harvest is

prior to receiving any income.

Religious Calendar. In addition to showing how employment patterns vary with

monthly harvesting season, it is important to rule out the presence of any religious signi�-

cance or activities associated with harvest which may explain the decline in violent activities.

Although Islamic religious festivities are common to all districts, its exact dates changes each

year. However, for the period under study in Iraq (2004-2009) and Pakistan (1988-2010) Ra-

madan always fell between August and October or August and January respectively, well

after the harvesting season in each case. Nonetheless, we make sure that harvesting does not

carry a local religious signi�cance that would explain the reduced violence and examine the

2008 Iraqi Time Use survey to examine whether the hours allocated to religious activities

vary by month. Figure 3 in the Appendix shows the coe�cients from a regression of hours

spent on religious activities on whether the individual is an agricultural worker or not. For

each month, there is no di�erence in religiosity among agricultural workers versus others.

However, we do observe a slight reduction in religious activities in June, the month when

about half of the districts experience harvest. This is consistent with the idea that the reduc-

tion in violence is unlikely to be driven by increased religiosity among agricultural workers.

In addition, in Table 10 and 11 of the online Appendix we estimate our baseline speci�-

cation using month of the year time e�ects (as opposed to only month with district by year

�xed e�ects separate) to account for any common factor a�ecting all districts in the same

25http://www.world-grain.com/Departments/Country-Focus/Iraq/Focus-on-Iraq.aspx?cck=1

28

month and year (e.g. religious festivals). Results are similar and more precisely estimated

for the IBC data � when compared to the baseline speci�cation � thus reducing any concern

that certain months may carry special signi�cance a�ecting con�ict intensity. The results

for WITS are less precisely estimated but similar in magnitude and sign.

4.2 The Pakistani Con�ict (1988-2010)

For the case of Pakistan we examine patterns of seasonal con�ict using the GTD Global

Terrorism Dataset (GTD) and the BFRS Political Violence Dataset (Bueno de Mesquita

et. al. 2015). These datasets categorize violents incident into whether it is conventional

(e.g. labor intensive) or asymmetric (e.g. less reliant on labor � IEDs, suicide bombs)

for Pakistan between 1988 and 2010.26 Given the constant presence of political violence

by di�erent militant groups in Pakistan, their high reliance on part-time forces, and the

importance of agriculture (in particular wheat) as a source of employment, we would expect

that seasonality play a role in con�ict intensity. For instance, Figure 4 of the online appendix,

shows how the peak of wheat harvesting in Pakistan occurs mostly in May, while planting

occurs mostly in October. This stands in contrast to Iraq's calendar, where most of the

harvesting occurs in June and the planting in December.

Table 2 below presents the results with the same speci�cation as before but using district-

level attacks in Pakistan between 1988 and 2010 in the GTD dataset. The �rst row shows

that the onset of harvest is associated with a reduction in the total number of attacks (column

1), the total number of asymmetric attacks by militants (column 2), conventional attacks

by militants (column3), and the monthly share of those killed (column 4). As noticed, the

onset of harvest is associated with a reduction in 0.4 percentage points in con�ict events,

yet, the higher average intensity of wheat production in these areas entail a higher average

e�ect ranging from 15 to 30% .

26More precisely, the authors of the BFRS dataset distinguish between militant, conventional and asym-metric attacks as follows �Militant attacks are those attributed to organized armed groups that use violencein pursuit of pre-de�ned political goals in ways that are: (a) planned; and (b) use weapons and tacticsattributed to sustained conventional or guerrilla warfare and not to spontaneous violence. Conventionalattacks by militants include direct conventional attacks on military, police, paramilitary, and intelligencetargets such that violence has the potential to be exchanged between the attackers and their targets. Asym-metric attacks include both terrorist attacks by militants, as well as militant attacks on military, police,paramilitary and intelligence targets that employ tactics that conventional forces do not, such as improvisedexplosive devices (IEDs).� (Bueno de Mesquita et. al. 2015: 17)

29

Table 2: Seasonal Labor and Violent Incidents in Pakistan

(1) (2) (3) (4)

GTDDV: Monthly % of... Total Attacks Asymm Laborint Tot Death

Harvim × Prodi -0.409*** -0.505*** -0.206 -0.369***(0.078) (0.103) (0.130) (0.103)

Plantcim × Prodci -0.092 -0.090 -0.141 -0.234**(0.131) (0.117) (0.176) (0.111)

Mean Harv Area 6.198 5.214 6.260 6.341Mean DV 8.333 8.333 8.333 8.333Avg E�ect -30.40 -31.62 -15.46 -28.06

Observations 7,704 5,484 4,056 5,964Clusters 111 95 100 105DistXYear FE Y Y Y YMonth FE Y Y Y YTemp & Precip Y Y Y Y

Clustered robust standard errors level in parentheses. Prodci is measured in hundredsq kilometers. DV in percentage points. *** p<0.01, ** p<0.05, * p<0.1

Cross-validation: BFRS Dataset (1988-2010). Additional evidence from the BFRS

shows some evidence in favor some seasonality of con�ict, though results are not as robust

as using GTD. Table 12 in the online appendix shows that while most coe�cients have

expected sign, only that of conventional attacks by militants is of statistical and economic

signi�cance, implying that the onset of harvest reduces these types of attacks 0.4 percentage

points, or 21% at means. However, this dataset may also mask signi�cant heterogeneity in

the extent to which rain-fed versus irrigation areas explain the results. As shown in Table

13 of the appendix, areas where wheat is rain-fed exhibit a stronger relationship between

con�ict intensity and seasonality relative to irrigated areas. In fact, the same relationship is

visible in the GTD dataset as shown in Table 14 of the online appendix. This suggests that

the distinction between rain-fed and irrigated is of greater signi�cant in Pakistan.

30

Robustness. In Table 15 of the online Appendix we run a regression interacting the

harvest indicator with a variable capturing whether the district is above or below the median

of wheat production intensity. We �nd that indeed most of the e�ect is driven by greater

wheat intensity cultivation in the GTD dataset. More importantly, in Table 16 we show how

the results are robust to including a month-year �xed e�ect (as opposed to a district-year

�xed e�ect with added month of the year �xed e�ect) which would account for a number

of alternative explanations such as variation in the strength of the overall con�ict, or levels

of religiosity between months, or changes in the state engagement, etc. Finally, Table 17

in the online appendix shows how most of the e�ect estimated above are driven by the

contemporaneous onset of harvest and not by its lags suggesting little anticipation e�ects.

4.3 Afghanistan (2004-2010)

After being overthrown by U.S. and U.K forces in 2001, the Taliban launched an insurgent

movement to regain power. Since then the insurgency has waged asymmetric warfare against

ISAF forces � the UN assistance force, later aided by NATO � as well as members of the

Afghan military and the government. Most of the Taliban recruits came from poor madras-

sas, motivated by local grievances, and participated only on a part-time basis due to their

work as farmers or laborers (Qazi, S. H. 2011: 10). Taliban cells were thus composed by

around ten to �fty part-time �ghters (Afsar, Samples, and Wood 2008: 65) who periodically

gather to launch attacks but then return to their laboring activities. Given their reliance on

part-time �ghters, it is likely that their availability and the intensity of the attacks will be

dictated by times of labor demand driven related to harvest.

One di�culty with the Afghan case is the presence of a highly lucrative opium trade which

has boomed with the Taliban presence. In fact, existing studies draw a connection between

31

con�ict and the incentives to cultivate opium (Lind et. al. 2014). While this connection

is interesting in its own right and an area for future research, it is a confounding factor in

our estimates, particularly because the conditions favoring wheat and opium production are

very similar, thus acting as substitutes. Since the harvest calendars overlap it makes it hard

to distinguish whether violence intensity is driven by wheat production or other dynamics

associated with illegal markets. The distinction is crucial given the huge di�erentials in value

created at harvest between wheat and poppy (�growing poppies is six times as pro�table as

growing wheat� UNODC 2010: 5) which may trigger appropriation incentives (or �rapacity

mechanisms�) dominating opportunity cost mechanisms as well as violence more generally,

such as in the Colombian case with coca production (Angrist and Kugler 2008). To account

for this, we depart from the baseline speci�cation in two ways: �rst, we limit the sample

to those districts where reported opium cultivation throughout the period is zero. However,

because this measure is naturally imprecise, we take advantage of the fact that most poppy

is cultivated in irrigated areas, while wheat is cultivated in both irrigated and rain-fed areas.

Hence focusing on the wheat calendar of rain-fed areas will better capture demand for labor

due to wheat cultivation as opposed to poppy.27 By examining only in rain-fed wheat in

non-opium provinces, we make sure that poppy cultivation is not present and unlikely to be

biasing our estimates.

Results from Table 3 below show indeed that focusing only in areas unlikely to be cul-

tivating poppy, the onset of harvest leads to a reduction in the intensity of attacks. The

coe�cient of -16 in column (1) suggests that the onset of harvest leads to a reduction on

average of 15 percentage points in the average of monthly attacks. However, given the av-

erage intensity of rain-fed wheat cultivation is only 0.092 hundred square kilometers, this

entails a 17% reduction in the average monthly share of attacks. In terms of types of attacks,

both asymmetric (bombs, �rearms) and labor intensive (attacks) are negatively related to

the onset of harvest. The same is true for a attacks initiated for overall casualties.

27Speci�cally we estimate: Attacksimt = αi+γit+β1(Harvimt×RainProdi)+β2(Harvimt×IrrigProdi)+ximt + eimt, where RainProdi and IrrigProdi are captured by the district �xed e�ect.

32

Table 3: Seasonal Labor and Violent Incidents in Afghanistan.

(1) (2) (3) (4)Provinces Below Median Opium Production

DV: Monthly % of... Total Attacks Asymm Attack Casualties

Harvcim ×RainProdci -14.962*** -14.862*** -20.941*** -21.225***(4.052) (4.197) (6.005) (6.963)

Plantcim ×RainProdci -8.805*** -8.843*** -8.451** -15.264***(2.801) (2.795) (3.945) (5.359)

Harvcim × IrrigProdci 1.129 0.285 -2.758 -1.241(2.924) (2.817) (2.076) (2.353)

Plantcim × IrrigProdci -2.343* -2.478* -2.625** -2.674*(1.252) (1.302) (1.281) (1.600)

Avg Harv Area 0.0928 0.0928 0.0896 0.0875Mean DV 8.333 8.333 8.333 8.333Avg E�ect -16.66 -16.55 -22.51 -22.28Observations 2,472 2,436 1,932 1,776Clusters 103 102 87 81District X Year FE Y Y Y YMonth FE Y Y Y YTemp& Precip Y Y Y YClustered robust standard errors at the district level in parentheses. Prodci is measured inhundred sq kilometers.*** p<0.01, ** p<0.05, * p<0.1

Robustness. Additional analysis in the online appendix shows that when combining

both irrigated and rain-fed in a single measure the result is less precisely estimated, po-

tentially driven by the fact that it is also capturing opium production (Table 18). Finally,

the inclusion of lags shows that for total attacks and casualties, the e�ect is driven by the

contemporaneous onset of harvest. However, in the case of total attacks and asymmetric

ones, we do observe a reduction in the intensity of attacks a month prior to the harvest,

suggesting some anticipation e�ects prior to the actual onset of harvest. in this case.

5 Conclusion

This paper has examined how seasonal variation in labor demand has a negative e�ect on

the intensity of violence. In Iraq, Pakistan, and Afghanistan, the number of attacks is

lower during harvest. Such a reduction in violent attacks ranges between 11 and 25% for

33

all cases when evaluated at the average share of monthly attacks and the average amount

of wheat cultivated in district. Results are robust to excluding regions that are not crop

producers, a wide array of �xed e�ect variables, and do not appear to be driven by alternative

explanations such as the weather, religious festivities, within-year variation in occupations,

or seasonal migration. Consistent with our interpretation that harvest a�ects local labor

markets and con�ict, we �nd that during these months agricultural workers tend to have

higher employment rates non-agricultural workers in Iraq. However, the way that attacks

are coded seems important: although there is some evidence of seasonality using the GTD in

Iraq and BFRS in Pakistan, estimated coe�cients are usually smaller and/or less precisely

estimated.

In terms of policy implications, care should be taken into interpreting our results for

the opportunity cost mechanism as evidence in favor of employment programs or permanent

forms of development aid. In theory, the problem is that those policy schemes may have

unintended consequences if highly persistent. For example, a permanent wage or employment

subsidy scheme may mean that households are wealthy enough to devote time to �ghting for

causes they care about, or are less likely to provide information to counter-insurgency forces.

Or, they may encourage people to �ght in order to capture the rents from these schemes.

Similarly, permanent changes in productivity (due to foreign or development aid) may have

a reduced e�ect of zero on violence, as �rst mentioned in Fearon (2008).

However, it might be possible to design more sophisticated policies which increase the

opportunity cost of violence without increasing either consumption or the value of winning.

For example, reducing food and energy subsidies (which are pervasive in regions prone to con-

�ict) and using the money for an employment subsidy would have little e�ect on the marginal

utility of consumption but would increase the incentive to work rather than �ght. Funding

employment schemes by local taxes would have a similar e�ect. Making employment subsi-

dies conditional on a successful counterinsurgency means they would not a�ect the value of

winning. These are just ideas: a thorough assessment would be an interesting area for future

research. An online appendix is available at https://sites.google.com/site/jennyguardado/.

34

References

[1] Afsar, S., Samples, C., and Wood, T. (2008). The Taliban: an organizational analysis.

Naval Postgraduate School Monterey CA. Department of National Security A�airs,

Chicago IL.

[2] Arjona AM and Kalyvas SN (2008) Rebelling against rebellion: Comparing insurgent

and counter-insurgent recruitment. Working paper. Yale, CT: Yale University.

[3] Bazzi, S., and Blattman, C. (2014). Economic shocks and con�ict: Evidence from com-

modity prices. American Economic Journal: Macroeconomics, 6(4), 1-38.

[4] Becker, G. 1968. �Crime and Punishment: An Economic Approach.� Journal of Political

Economy 76(2):169-217.

[5] Berman, E, J N. Shapiro and J H. Felter. 2011a. �Can Hearts and Minds be Bought? The

Economics of Counterinsurgency in Iraq.� Journal of Political Economy 119:766-819.

[6] Berman, E, M Callen, J H. Felter and J N. Shapiro. 2011b. �Do Working Men Rebel?

Insurgency and Unemployment in Afghanistan, Iraq, and the Philippines.� Journal of

Con�ict Resolution 55(4)

[7] Blattman C and Miguel E (2010) Civil war. Journal of Economic Literature 48(1): 3�57.

[8] Blattman C and Ralston L (2015) Generating employment in poor and fragile states:

Evidence from labor market and entrepreneurship programs. Mimeo.

[9] Bertrand M., E. Du�o and S. Mullianathan (2004) �Hoe much should we trust

di�erences-in-di�erences estimates?� The Quarterly Journal of Economics 119 (1)

[10] Brückner, Markus, and Antonio Ciccone. "International Commodity Prices, Growth

and the Outbreak of Civil War in Sub-Saharan Africa." The Economic Journal 120, no.

544 (2010): 519-534.

[11] Bueno de Mesquita, E. (2013). Rebel tactics. Journal of Political Economy, 121(2),

323-357.

[12] Bueno de Mesquita, et. al. (2014). Measuring political violence in Pakistan: Insights

from the BFRS Dataset. Con�ict Management and Peace Science.

35

[13] Burke, M. B., Miguel, E., Satyanath, S., Dykema, J. A., & Lobell, D. B. (2009). Warm-

ing increases the risk of civil war in Africa. Proceedings of the National Academy of

Sciences, 106(49), 20670-20674.

[14] Ciccone, A. (2011). Economic shocks and civil con�ict: A comment. American Economic

Journal: Applied Economics, 3(4), 215-227.

[15] Chassang S and Padró i Miquel G (2009) Econmic shocks and civil war. Quarterly

Journal of Political Science 4(3): 211�228.

[16] Collier P and Hoe�er A (2004) Greed and grievance in civil war. Oxford Economic

Papers 56(4): 563�595.

[17] Crost, B., Felter, J., & Johnston, P. (2014). Aid under �re: Development projects and

civil con�ict. The American Economic Review, 104(6), 1833-1856.

[18] Dal Bo, Ernesto, and Dal Bo, Pedro. (2011). �Workers, warriors, and criminals: social

con�ict in general equilibrium.�Journal of the European Economic Association, 9(4),

646-677.

[19] Dube, Oeindrila, and Juan F. Vargas. "Commodity price shocks and civil con�ict: Ev-

idence from Colombia." The Review of Economic Studies 80, no. 4 (2013): 1384-1421.

[20] Fearon J. (2005) Primary commodity exports and civil war. Journal of Con�ict Reso-

lution 49(4): 483�507.

[21] Fearon J. and Laitin D. (2003) Ethnicity, insurgency, and civil war. American Political

Science Review 97(1): 75�90.

[22] Fearon, J. (2008). Economic development, insurgency, and civil war. Institutions and

economic performance, 292, 328.

[23] Figes, Orlando (1996). A People's Tragedy: The Russian Revolution 1891�1924. Pimlico,

London.

[24] Gar�nkel, Michelle R. 1990. �Arming as a Strategic Investment in a Cooperative Equi-

librium.� American Economic Review, 80(1): 50�68.

[25] Gar�nkel, Michelle R., and Stergios Skaperdas. 2007. �Economics of Con�ict: An

Overview.� In Handbook of Defense Economics, Volume 2, Defense in a Globalized

36

World, ed. Todd Sandler, and Keith Hartley, 649�710. Amsterdam and Oxford: Else-

vier, North-Holland.

[26] Giu�re, K. A. (1997). First in Flight: Desertion as Politics in the North Carolina Con-

federate Army. Social Science History, 245-263.

[27] Grossman, Herschel I. 1991. �A General Equilibrium Model of Insurrections.� American

Economic Review, 81(4): 912�21.

[28] Guardado, J. 2015. �Land Tenure, Price Shocks and Insurgency: Evidence from Peru

and Colombia�. Mimeo.

[29] Haavelmo, Trygve. 1954. A Study in the Theory of Economic Evolution. Amsterdam:

North-Holland.

[30] Hirshleifer, Jack. 1988. �The Analytics of Continuing Con�ict.� Synthese, 76(2): 201�33.

[31] Hirshleifer, Jack. 1989. �Con�ict and Rent-Seeking Success Functions: Ratio vs. Di�er-

ence Models of Relative Success.� Public Choice, 63(2): 101�12.

[32] Hodler, Roland, and Paul A. Raschky. "Economic shocks and civil con�ict at the regional

level." Economics Letters 124, no. 3 (2014): 530-533.

[33] Horowitz, Donald L. 1985. Ethnic Groups in Con�ict. Berkeley and London: University

of California Press.

[34] Hsiang, S. M., Burke, M., and Miguel, E. (2013). Quantifying the in�uence of climate

on human con�ict. Science, 341(6151), 1235367.

[35] Lind, J. T., Moene, K. O., & Willumsen, F. (2014). Opium for the masses? Con�ict-

induced narcotics production in Afghanistan. Review of Economics and Statistics, 96(5),

949-966.

[36] McClintock, Cynthia. 1998. Revolutionary Movements in Latin America: El Salvador's

FMLN and Peru's Shining Path. Washington, DC: United States Institute of Peace.

[37] Miguel, Edward, Shanker Satyanath and Ernest Sergenti. 2004. �Economic Shocks and

Civil Con�ict: An Instrumental Variables Approach.� Journal of Political Economy

112:725-753

37

[38] Nunn, N., & Qian, N. (2014). US food aid and civil con�ict. The American Economic

Review, 104(6), 1630-1666.

[39] Paxson, C. H. (1992). Using weather variability to estimate the response of savings to

transitory income in Thailand. The American Economic Review, 15-33.

[40] Qazi, S. H. (2011). The neo-taliban, counterinsurgency & the american endgame in

afghanistan. Institute for Social Policy and Understanding.

[41] Sarsons, Heather. "Rainfall and con�ict." In Manuscript. http://www. econ. yale.

edu/conference/neudc11/papers/paper_199. pdf. 2011.

[42] Skaperdas, Stergios. 1992. �Cooperation, Con�ict, and Power in the Absence of Property

Rights.� American Economic Review, 82(4): 720�39.

[43] Watkin, K., & Norris, A. J. (2012). Non-international Armed Con�ict in the Twenty-�rst

Century. Government Printing O�ce.

[44] Wigle, J. (2010). Introducing the worldwide incidents tracking system (WITS). Per-

spectives on Terrorism, 4(1).

[45] Willmott, C. J. and K. Matsuura (2001) Terrestrial Air Temperature

and Precipitation: Monthly and Annual Time Series (1950 - 1999),

http://climate.geog.udel.edu/~climate/html_pages/README.ghcn_ts2.html.

[46] Wood, E. J. (2003). Insurgent collective action and civil war in El Salvador. Cambridge

University Press.

38